Key points are not available for this paper at this time.
State-of-charge (SOC) estimation is critical for reliable operation of Li-ion batteries (LIBs). However, the distinct electrochemical characteristics coupled with harsh low-temperature environments make a single estimator struggle to robustly estimate the volatile SOC of multi-type LIBs. To address these issues, this article proposes a hard-soft hybrid prompt learning method to unleash the potential of a pretrained large language model (LLM) for SOC estimation. A textual encoder is introduced to convert LIB measurements into hard text prompts for language modeling, naturally eliciting the pretrained LLM to capture the intra-relations of measured values over time and their inter-relations with contextual semantics for accurate estimates. A side adapter network is constructed to reparameterize model adaptation towards different LIB tasks into optimizations within a low-dimensional subspace, strengthening the estimation generalization of the pretrained LLM in a parameter-efficient manner. A knowledge infusion mechanism is designed to encapsulate task-specific information as soft prompt vectors for model integration along forward propagation, dynamically conditioning the hidden states inside the pretrained LLM to enhance the estimation robustness against SOC volatilities. Extensive experiments verify that the hybrid prompt-driven LLM can simultaneously perform estimations for multi-type LIBs under diverse operations and sub-zero temperatures with superior accuracy, generalization, and robustness.
Bian et al. (Mon,) studied this question.